Skip to content

[core] reuse AttentionMixin for compatible classes #12463

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 10 commits into
base: main
Choose a base branch
from
Open

[core] reuse AttentionMixin for compatible classes #12463

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
06d4fb5
remove attn_processors property
sayakpaul Oct 10, 2025
2c9b7c9
more
sayakpaul Oct 10, 2025
f4a2bde
up
sayakpaul Oct 10, 2025
c04b016
up more.
sayakpaul Oct 10, 2025
339f9e5
Merge branch 'main' into reuse-attn-mixin
sayakpaul Oct 11, 2025
6dbee1e
up
sayakpaul Oct 11, 2025
95fcb53
add AttentionMixin to AuraFlow.
sayakpaul Oct 14, 2025
98d2123
Merge branch 'main' into reuse-attn-mixin
sayakpaul Oct 14, 2025
fabc15d
up
sayakpaul Oct 14, 2025
20dd719
Merge branch 'main' into reuse-attn-mixin
sayakpaul Oct 15, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
66 changes: 3 additions & 63 deletions src/diffusers/models/autoencoders/autoencoder_kl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Dict, Optional, Tuple, Union
from typing import Optional, Tuple, Union

import torch
import torch.nn as nn
Expand All @@ -21,11 +21,11 @@
from ...loaders.single_file_model import FromOriginalModelMixin
from ...utils import deprecate
from ...utils.accelerate_utils import apply_forward_hook
from ..attention import AttentionMixin
from ..attention_processor import (
ADDED_KV_ATTENTION_PROCESSORS,
CROSS_ATTENTION_PROCESSORS,
Attention,
AttentionProcessor,
AttnAddedKVProcessor,
AttnProcessor,
FusedAttnProcessor2_0,
Expand All @@ -35,7 +35,7 @@
from .vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder


class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalModelMixin, PeftAdapterMixin):
class AutoencoderKL(ModelMixin, AttentionMixin, ConfigMixin, FromOriginalModelMixin, PeftAdapterMixin):
r"""
A VAE model with KL loss for encoding images into latents and decoding latent representations into images.

Expand Down Expand Up @@ -167,66 +167,6 @@ def disable_slicing(self):
"""
self.use_slicing = False

@property
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
def attn_processors(self) -> Dict[str, AttentionProcessor]:
r"""
Returns:
`dict` of attention processors: A dictionary containing all attention processors used in the model with
indexed by its weight name.
"""
# set recursively
processors = {}

def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
if hasattr(module, "get_processor"):
processors[f"{name}.processor"] = module.get_processor()

for sub_name, child in module.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

return processors

for name, module in self.named_children():
fn_recursive_add_processors(name, module, processors)

return processors

# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
r"""
Sets the attention processor to use to compute attention.

Parameters:
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
The instantiated processor class or a dictionary of processor classes that will be set as the processor
for **all** `Attention` layers.

If `processor` is a dict, the key needs to define the path to the corresponding cross attention
processor. This is strongly recommended when setting trainable attention processors.

"""
count = len(self.attn_processors.keys())

if isinstance(processor, dict) and len(processor) != count:
raise ValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
)

def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
if hasattr(module, "set_processor"):
if not isinstance(processor, dict):
module.set_processor(processor)
else:
module.set_processor(processor.pop(f"{name}.processor"))

for sub_name, child in module.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

for name, module in self.named_children():
fn_recursive_attn_processor(name, module, processor)

# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Perhaps it's out of the scope for this PR, but I see that a lot of models additionally have a set_default_attn_processor method, usually # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor. Do you think it makes sense to add this method to AttentionMixin?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

IMO, not yet since AttentionMixin is fairly agnostic to the model-type but set_default_attn_processor relies on some custom attention processor types. For UNet2DConditionModel, we have:

diffusers/src/diffusers/models/unets/unet_2d_condition.py

Lines 762 to 769 in fa468c5

if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnAddedKVProcessor()
elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnProcessor()
else:
raise ValueError(
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
)

However, for AutoencoderKL Temporal Decoder:

diffusers/src/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py

Lines 269 to 274 in fa468c5

if all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
processor = AttnProcessor()
else:
raise ValueError(
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
)

I'd be down to the refactoring, though. Cc: @DN6

def set_default_attn_processor(self):
"""
Expand Down
67 changes: 4 additions & 63 deletions src/diffusers/models/autoencoders/autoencoder_kl_temporal_decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,14 +12,15 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import itertools
from typing import Dict, Optional, Tuple, Union
from typing import Optional, Tuple, Union

import torch
import torch.nn as nn

from ...configuration_utils import ConfigMixin, register_to_config
from ...utils.accelerate_utils import apply_forward_hook
from ..attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor
from ..attention import AttentionMixin
from ..attention_processor import CROSS_ATTENTION_PROCESSORS, AttnProcessor
from ..modeling_outputs import AutoencoderKLOutput
from ..modeling_utils import ModelMixin
from ..unets.unet_3d_blocks import MidBlockTemporalDecoder, UpBlockTemporalDecoder
Expand Down Expand Up @@ -135,7 +136,7 @@ def forward(
return sample


class AutoencoderKLTemporalDecoder(ModelMixin, ConfigMixin):
class AutoencoderKLTemporalDecoder(ModelMixin, AttentionMixin, ConfigMixin):
r"""
A VAE model with KL loss for encoding images into latents and decoding latent representations into images.

Expand Down Expand Up @@ -202,66 +203,6 @@ def __init__(

self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)

@property
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
def attn_processors(self) -> Dict[str, AttentionProcessor]:
r"""
Returns:
`dict` of attention processors: A dictionary containing all attention processors used in the model with
indexed by its weight name.
"""
# set recursively
processors = {}

def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
if hasattr(module, "get_processor"):
processors[f"{name}.processor"] = module.get_processor()

for sub_name, child in module.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

return processors

for name, module in self.named_children():
fn_recursive_add_processors(name, module, processors)

return processors

# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
r"""
Sets the attention processor to use to compute attention.

Parameters:
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
The instantiated processor class or a dictionary of processor classes that will be set as the processor
for **all** `Attention` layers.

If `processor` is a dict, the key needs to define the path to the corresponding cross attention
processor. This is strongly recommended when setting trainable attention processors.

"""
count = len(self.attn_processors.keys())

if isinstance(processor, dict) and len(processor) != count:
raise ValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
)

def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
if hasattr(module, "set_processor"):
if not isinstance(processor, dict):
module.set_processor(processor)
else:
module.set_processor(processor.pop(f"{name}.processor"))

for sub_name, child in module.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

for name, module in self.named_children():
fn_recursive_attn_processor(name, module, processor)

def set_default_attn_processor(self):
"""
Disables custom attention processors and sets the default attention implementation.
Expand Down
66 changes: 3 additions & 63 deletions src/diffusers/models/autoencoders/consistency_decoder_vae.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from dataclasses import dataclass
from typing import Dict, Optional, Tuple, Union
from typing import Optional, Tuple, Union

import torch
import torch.nn.functional as F
Expand All @@ -23,10 +23,10 @@
from ...utils import BaseOutput
from ...utils.accelerate_utils import apply_forward_hook
from ...utils.torch_utils import randn_tensor
from ..attention import AttentionMixin
from ..attention_processor import (
ADDED_KV_ATTENTION_PROCESSORS,
CROSS_ATTENTION_PROCESSORS,
AttentionProcessor,
AttnAddedKVProcessor,
AttnProcessor,
)
Expand All @@ -49,7 +49,7 @@ class ConsistencyDecoderVAEOutput(BaseOutput):
latent_dist: "DiagonalGaussianDistribution"


class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
class ConsistencyDecoderVAE(ModelMixin, AttentionMixin, ConfigMixin):
r"""
The consistency decoder used with DALL-E 3.

Expand Down Expand Up @@ -200,66 +200,6 @@ def disable_slicing(self):
"""
self.use_slicing = False

@property
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
def attn_processors(self) -> Dict[str, AttentionProcessor]:
r"""
Returns:
`dict` of attention processors: A dictionary containing all attention processors used in the model with
indexed by its weight name.
"""
# set recursively
processors = {}

def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
if hasattr(module, "get_processor"):
processors[f"{name}.processor"] = module.get_processor()

for sub_name, child in module.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

return processors

for name, module in self.named_children():
fn_recursive_add_processors(name, module, processors)

return processors

# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
r"""
Sets the attention processor to use to compute attention.

Parameters:
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
The instantiated processor class or a dictionary of processor classes that will be set as the processor
for **all** `Attention` layers.

If `processor` is a dict, the key needs to define the path to the corresponding cross attention
processor. This is strongly recommended when setting trainable attention processors.

"""
count = len(self.attn_processors.keys())

if isinstance(processor, dict) and len(processor) != count:
raise ValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
)

def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
if hasattr(module, "set_processor"):
if not isinstance(processor, dict):
module.set_processor(processor)
else:
module.set_processor(processor.pop(f"{name}.processor"))

for sub_name, child in module.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

for name, module in self.named_children():
fn_recursive_attn_processor(name, module, processor)

# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
def set_default_attn_processor(self):
"""
Expand Down
64 changes: 2 additions & 62 deletions src/diffusers/models/controlnets/controlnet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,10 +21,10 @@
from ...configuration_utils import ConfigMixin, register_to_config
from ...loaders.single_file_model import FromOriginalModelMixin
from ...utils import BaseOutput, logging
from ..attention import AttentionMixin
from ..attention_processor import (
ADDED_KV_ATTENTION_PROCESSORS,
CROSS_ATTENTION_PROCESSORS,
AttentionProcessor,
AttnAddedKVProcessor,
AttnProcessor,
)
Expand Down Expand Up @@ -106,7 +106,7 @@ def forward(self, conditioning):
return embedding


class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
class ControlNetModel(ModelMixin, AttentionMixin, ConfigMixin, FromOriginalModelMixin):
"""
A ControlNet model.

Expand Down Expand Up @@ -515,66 +515,6 @@ def from_unet(

return controlnet

@property
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
def attn_processors(self) -> Dict[str, AttentionProcessor]:
r"""
Returns:
`dict` of attention processors: A dictionary containing all attention processors used in the model with
indexed by its weight name.
"""
# set recursively
processors = {}

def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
if hasattr(module, "get_processor"):
processors[f"{name}.processor"] = module.get_processor()

for sub_name, child in module.named_children():
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

return processors

for name, module in self.named_children():
fn_recursive_add_processors(name, module, processors)

return processors

# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
r"""
Sets the attention processor to use to compute attention.

Parameters:
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
The instantiated processor class or a dictionary of processor classes that will be set as the processor
for **all** `Attention` layers.

If `processor` is a dict, the key needs to define the path to the corresponding cross attention
processor. This is strongly recommended when setting trainable attention processors.

"""
count = len(self.attn_processors.keys())

if isinstance(processor, dict) and len(processor) != count:
raise ValueError(
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
)

def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
if hasattr(module, "set_processor"):
if not isinstance(processor, dict):
module.set_processor(processor)
else:
module.set_processor(processor.pop(f"{name}.processor"))

for sub_name, child in module.named_children():
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

for name, module in self.named_children():
fn_recursive_attn_processor(name, module, processor)

# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
def set_default_attn_processor(self):
"""
Expand Down
Loading
Loading